Khiliad Legal
Back to Insights

SRA Guidance on AI: What Solicitors Need to Know and Do

A plain English guide to the SRA's position on AI: the August 2026 warning notice, the June 2026 supervision guidance, the competence consultation, the court cases behind them, and a practical checklist.

Naz Keceli15 minute read17 November 2026

The SRA hasn't written a new rule for AI, and it says it doesn't need one. In a warning notice published on 17 August 2026, it made a simple point: every duty a solicitor already has applies in full when AI is involved. If an AI tool invents a case and it ends up in your skeleton argument, it's still your skeleton argument. If a trainee pastes a client's medical records into a free chatbot, it's still your firm's confidentiality breach. And if you signed off the work without knowing how it was produced, that's a supervision failure.

None of that is new law. What's new is that the regulator has said it out loud, tied it to specific paragraphs of the Codes, and made clear it will act. From September 2027, competent use of AI is also expected to become part of the Statement of Solicitor Competence, the standard every solicitor is measured against.

This guide walks through what the SRA has actually said, the court decisions that pushed it there, what's coming next, and what your firm should be able to show if the SRA, a court or your insurer ever asks how you use AI.

This is general information about the regulatory position, not legal advice. Read the SRA's own documents, linked at the end, before relying on it.

How we got here

It's worth knowing the order things happened in, because each step built on the one before.

In June 2025 the Divisional Court gave judgment in Ayinde v Haringey, two cases where lawyers had put fictitious authorities before the court. Dame Victoria Sharp, President of the King's Bench Division, said lawyers who use AI for research have "a professional duty therefore to check the accuracy of such research". The lawyers were referred to their regulators.

In November 2025 the Upper Tribunal dealt with fake citations from two more firms, in a decision now reported as [2026] UKUT 81 (IAC). It went further than Ayinde, on supervision and on confidentiality, and we come back to it below.

By the summer of 2026 the SRA had received 42 reports of possible AI misuse in a single year, covering inaccurate citations, supervision and confidentiality. In June it added a section on AI to its supervision guidance. In August came the warning notice. And in October it opened a consultation on writing AI into the competence statement, which closes on 3 December 2026.

All the while, the numbers kept climbing. An independent database kept by researcher Damien Charlotin listed 76 UK court decisions involving fabricated AI output by 19 October 2026, out of 2,202 worldwide.

The SRA is careful to say it isn't against AI. When it published the notice, its Executive Director of Strategy and Policy, Aileen Armstrong, said AI "is already benefiting consumers and legal professionals", but that "the use of AI does not change the professional standards expected of solicitors and law firms."

That's the whole position in one sentence. The problem isn't the technology. It's work nobody checked, client data put in the wrong place, and supervisors who didn't know how the work in front of them was made.

Our report, The impact of AI in the legal sector, sets out the wider picture, including how far AI use by individual lawyers has run ahead of what firms have put in place.

What the SRA warning notice on AI says

The notice applies to everyone the SRA regulates, in-house solicitors included. It starts with two concerns.

The first is inaccurate material, especially in court. AI tools can produce cases, citations and quotations that look entirely convincing and simply don't exist. The notice is blunt about who carries the responsibility. AI has "no separate legal personality", it says, and relying on what a tool produced is not a defence.

The second is confidentiality. The SRA's worry is that firms haven't thought through what happens to client information once someone types or uploads it into an AI tool, free or paid. It expects contractual, technical and organisational safeguards to be in place before any client information goes in. In practice that means a secure environment, no unauthorised access, no use of your data to train the model unless you've agreed to it, and nothing kept for longer than it needs to be.

From there, the notice sets out what it expects. Solicitors stay accountable for AI assisted work, and they have to supervise the junior and non-authorised staff who use it. Firms need governance, systems and controls that cover the risks of AI, scaled to the size of the firm and the work it does. Every authority cited has to be genuine, relevant and checkable from its citation. Human oversight and professional judgement apply throughout, in a proportionate, risk based way. In-house solicitors get a specific mention too, which we cover near the end.

The closing warning is the part that should get a partner's attention. Failing to have proper regard to the notice puts firms and individuals at risk of disciplinary action. The SRA says it "can and will act" where the rules are broken, and putting false material before a court may also amount to contempt.

Why AI makes things up, and why lawyers are so exposed

It helps to understand why this keeps happening, because it shapes how you check.

A general purpose AI tool doesn't look things up the way a legal database does. It writes the most likely next words, based on patterns in the text it learned from. Legal citations follow a very regular pattern: party names, a year in square brackets, a court, a number. So a tool can produce something that looks exactly like a real citation, attached to a proposition that sounds exactly like something a judge might say. Nothing inside the tool knows whether it's real.

That's why "it looked right" is no defence, and why the check has to be against the source itself. A proper check means opening the judgment, on the National Archives' Find Case Law service, BAILII or your subscription database, finding the paragraph being relied on, reading it in context, and making sure the case hasn't been overturned or distinguished since. Newer tools that search the web or a legal database are better at this, but they can still misquote a real case or pin a point to the wrong paragraph. The rule doesn't change: if you're relying on it, someone has read it.

The risk isn't only in your own work, either. Of the 2,202 decisions in the Charlotin database, 1,267 involved litigants in person and 872 involved lawyers. If you litigate against unrepresented parties, you'll increasingly see invented authorities from the other side, and they need checking as carefully as your own.

The rules the notice relies on

The notice doesn't create new duties. It ties AI to rules that already exist. The SRA Principles it cites are 1 (the rule of law and the proper administration of justice), 2 (public trust and confidence), 4 (honesty), 5 (integrity) and 7 (each client's best interests). These are the Code paragraphs most firms will recognise, paraphrased in plain English.

Code of Conduct for SolicitorsIn plain EnglishWhere AI comes in
1.4Do not mislead clients, the court or othersAn invented case or quotation misleads, whoever wrote it
2.4Only put forward submissions that are properly arguableAn argument built on authorities that do not exist is not arguable
2.7Draw the court's attention to relevant cases and provisionsCiting real authorities, accurately, is part of the duty
3.2Provide a competent serviceUsing a tool you do not understand, or not checking it, is not competent
3.5Stay accountable for work done by people you superviseYou answer for a trainee's AI assisted draft as if you wrote it
3.6Make sure the people you manage are competent and up to dateStaff using AI need training on how to use it safely
6.3Keep clients' affairs confidentialClient information in an unsafe tool can breach confidentiality
7.2Be able to justify your decisions and show complianceYou need to be able to show what was checked, and how

The notice also cites the Code of Conduct for Firms, including 2.1 (governance, systems and controls), 4.2 to 4.4 (competent service, competent staff and effective supervision), 6.3 (confidentiality) and 8.1 (managers' responsibility for compliance), as well as the COLP's duties.

Put simply, this is a firm problem as well as an individual one. A fake citation is the fee earner's mistake. The absence of any system that would have caught it belongs to the firm.

What the courts have said

The warning notice leans heavily on three recent judgments. Each one teaches something slightly different.

Ayinde v Haringey (June 2025): you can't hand the check to someone else

The Divisional Court dealt with two cases where fictitious authorities had reached the court. It said lawyers who use AI for research have "a professional duty therefore to check the accuracy of such research", and that "a lawyer is not entitled to rely on their lay client for the accuracy of citations."

That second line matters more than it first appears. It closes off the argument that someone else supplied the material. Whoever puts it before the court owns it.

The lawyers involved were referred to their regulators, and the court sent its judgment to the Law Society, the Bar Council and the Inns of Court asking them to act. The court wanted the profession to treat this as a conduct issue, not a one off embarrassment.

UK v SSHD [2026] UKUT 81 (IAC): supervisors are judged more harshly

The Upper Tribunal dealt with fake citations from two firms, and made three points every firm should take seriously.

The first is about supervision. A professional who delegates work stays responsible for it, and the tribunal said a supervisor who fails to spot a junior's false citations is likely to be more culpable than someone who fails to check their own work. That turns the usual instinct on its head. It's tempting to see the junior who used the tool as the one at fault. The tribunal sees the supervisor who let it through as the bigger problem.

The second is the statement of truth. The Upper Tribunal's judicial review claim form now asks the representative to confirm that every authority cited exists, can be found from its citation, and supports the point it's cited for. Signing that while citing false authorities should ordinarily lead to a regulatory referral. It's a small change on a form, but it means the check is now something you personally put your name to. Notably, one of the two solicitors in the case avoided a referral because he had reported himself first.

The third is confidentiality. The tribunal said putting client documents into an open AI tool such as ChatGPT places them in the public domain, breaching confidentiality and waiving privilege, and that such conduct should be referred to the ICO. The warning is right. Some of the technical framing is less precise, as we explain in our analysis of UKUT 81 and our technical note on AI tools and client data. What protects client data is the plan and the contract your firm is on, not the brand of the tool.

Cork v Smith [2026] EWHC 1199 (Ch): size is no protection

In the High Court, a junior solicitor at a large firm used AI to draft two letters to the court that misstated the Insolvency Rules. The judge said the junior "seems to have almost entirely outsourced the thinking process to the program" and that "there appears to have been a failure to supervise". The firm referred itself to the SRA.

It's easy to assume this only happens at small, stretched firms. Cork v Smith shows otherwise. A large firm with resources and procedures still had a junior leaning on a tool and a supervisor who didn't catch it. What matters is whether the controls work on an ordinary busy afternoon, not whether they exist on paper.

Supervising AI assisted work

Before the warning notice, in June 2026, the SRA added a section on AI to its effective supervision guidance. It asks firms to "consider what effective supervision of AI assisted or AI generated work looks like in practice", including making sure AI outputs get "appropriate human review", and that an authorised person "retains ultimate responsibility" for legal services delivered with AI.

So what does that look like on a real file? Traditional supervision looks at the work itself. Is the advice right, is the letter well drafted, are the authorities on point? Supervising AI assisted work adds questions about how the work was made. Was AI used, and for what? What did the person check, and against what?

You don't need a new form for that. A line on the file note does the job: first draft prepared with the firm's approved tool, authorities checked against the source by a named person on a given date. It takes ten seconds to write. If the SRA or a court ever asks, it's the difference between saying you had a process and showing that you followed it.

Smaller firms have an extra problem. In a firm of five, the COLP is often also a fee earner, the COFA and whoever sorts out the IT. The supervision duty doesn't shrink with the firm, so the process has to be light enough that a busy partner actually follows it.

What's coming next: AI in the Statement of Solicitor Competence

On 8 October 2026 the SRA opened a consultation on the Statement of Solicitor Competence. It proposes writing technology and AI into four places. Maintaining competence would include keeping up with developments in technology, including AI. Legal research would include checking the outputs of technology, AI included, against primary sources. Drafting would include making appropriate use of technology. And a new standalone section on supervision would include overseeing how technology is used.

The consultation closes at 23:59 on 3 December 2026. The SRA plans to publish the final statement in April 2027, with the changes taking effect from September 2027.

If it goes through as drafted, competent practice will include competent use of AI by definition. That matters even for firms that have decided not to use AI at all. A solicitor who has never used these tools will struggle to supervise someone who has, and from 2027, supervising it competently is part of the job. Banning AI doesn't avoid the duty. It leaves the firm less able to meet it, while people quietly use free tools anyway. That gap has a name now: shadow AI.

Does the SRA say you must tell clients you use AI?

No. There is no SRA rule requiring you to disclose AI use to clients. The Law Society's guidance, Generative AI: the essentials, suggests deciding with clients whether and how AI tools are used, and communicating clearly when they are used on a matter where appropriate. It also points to the engagement letter, which may govern how client data can be used.

Most firms now take one of three approaches: a line in the terms of business, telling clients when AI is used in a material way on their matter, or answering honestly if asked and following any instruction not to use it. Whichever you choose, write it down in your AI policy and apply it consistently.

Does the SRA ban ChatGPT or any other tool?

No. The SRA doesn't approve or ban products. What it expects is that client information only goes into a tool with the right safeguards: a contract that rules out training on your data, a secure environment, sensible retention and controlled access.

That usually rules out free and personal accounts for client work, whatever the brand. Business plans such as ChatGPT Business, Claude Team and Microsoft 365 Copilot can meet the standard when they're set up properly. Our guides to whether ChatGPT is safe for client data, Claude for law firms and Microsoft Copilot for law firms explain the differences.

If a fake citation reaches the court

Picture how it usually unfolds. The other side, or the judge, can't find a case cited in your skeleton argument. The court asks for an explanation. Depending on how it went wrong and how it's handled, that can lead to a wasted costs order, a referral to the SRA and, in serious cases, contempt proceedings.

How the firm responds in the first few days matters a great deal. Tell the court as soon as the error is found, correct the record, and work out quickly whether anything else on the file came from the same source. Consider reporting to the SRA yourself rather than waiting. In UKUT 81, that was the difference between a referral and no referral. And if client information went into an unsafe tool along the way, treat it as a possible data breach, because a reportable breach has to go to the ICO within 72 hours.

If the SRA does get involved, its first question is unlikely to be "did you use AI?" It's far more likely to be "what did you have in place?" That's when it helps to show, quickly and on paper, that the firm took this seriously before anything went wrong. These are the things a well run firm should be able to put on the table.

What to have in placeWhy it mattersWhere it comes from
A list of approved AI tools, on business plansShows client data only goes where safeguards existWarning notice, Code 6.3
A written AI policy that everyone knowsShows governance, systems and controlsCode for Firms 2.1, warning notice
A rule that every authority is checked against the sourceShows verification is routine, not optionalWarning notice, Ayinde, Code 1.4 and 2.4
A supervision step that records what was checkedShows accountability for AI assisted workCode 3.5, supervision guidance, UKUT 81
Training records for everyone using AIShows staff are competent and up to dateCode 3.6, competence consultation
Data processing agreements and impact assessmentsShows data protection was considered firstWarning notice, UK GDPR
A clear way to report mistakesShows problems are caught and fixed earlyCode 7.2, data breach duties

A practical checklist for your firm

None of this needs a big project. Most firms can have the basics in place within a month, and the order matters, because each step makes the next one easier.

  1. Find out what's already happening. Ask every team, without blame, which AI tools they use and for what. In most firms the honest answer includes free tools on personal accounts that nobody approved. That isn't a disaster. It's your starting point.
  2. Approve at least one safe tool. Pick a business plan, check the data terms, sign the data processing agreement and set up access properly. People use free tools because nobody gave them a better option, so giving them one is the fastest way to bring that use into the open.
  3. Write a short policy. Cover which tools are approved, what can and can't be shared, how output is checked, who is responsible and who to ask. One or two pages that people actually read beats thirty that nobody does. Our free AI policy template for law firms is a starting point.
  4. Make checking authorities non-negotiable. Every case, statute and quotation is checked against the original source before anyone relies on it. For court work, assume you'll be asked to confirm it, because increasingly you will.
  5. Add one line to the sign off. For AI assisted work, the supervisor notes what was checked and how. It takes seconds and it's your evidence later.
  6. Train everyone, and keep a record. Include support staff, who often use these tools as much as fee earners. A dated training record per person will also help with the competence statement from September 2027. Our guide to AI training for solicitors covers what good training looks like.
  7. Give it an owner. Usually the COLP or a named partner, who keeps the policy and the tool list up to date.
  8. Review it regularly. At least every six months, and whenever the SRA or the courts say something new. They've said something new roughly every quarter for the past year.
  9. Consider responding to the consultation. It closes on 3 December 2026 and takes about an hour. The final statement will apply to your firm whether you respond or not, so it's worth having a say.

The warning notice applies to in-house solicitors too, and calls out their position specifically. The tools available to an in-house team are often chosen by the wider business, for business reasons, and nobody in procurement was thinking about privilege when they signed the contract. In-house solicitors need to check those tools meet their professional duties on confidentiality and privilege, and be ready to say so when the business's enthusiasm for AI and their own obligations pull in different directions.

Sources

How we help

Khiliad Legal helps firms put the SRA's expectations into practice. Our AI Day sets up an approved AI tool with the right data terms, writes your AI policy with the partners, and trains the people who will lead it, all in one day. If you want to know where your firm stands first, our AI Readiness Assessment looks at how AI is already used across the firm, measured against the SRA or your own regulator.

Want to know how your firm measures up? Book a short call with us.

Frequently asked questions

Has the SRA issued guidance on AI?

Yes. The main document is the SRA's warning notice on misuse of AI, published on 17 August 2026. The SRA also added an AI section to its effective supervision guidance in June 2026, and in October 2026 it opened a consultation on adding AI to the Statement of Solicitor Competence.

Is it a breach of the SRA rules to use ChatGPT?

Not in itself. The risk comes from how it's used: putting client information into a tool without proper safeguards, or relying on output that nobody checked. Free and personal accounts are rarely suitable for client work. A properly set up business plan can be.

Is relying on AI a defence if something goes wrong?

No. The SRA's warning notice says AI has no separate legal personality, and the courts have said the same. The solicitor responsible for the work is responsible for every error in it.

Who is responsible when a trainee uses AI?

Both the trainee and the supervisor. Code paragraph 3.5 keeps supervisors accountable for work done through them, and in UKUT 81 the Upper Tribunal said a supervisor who misses a junior's fake citations is likely more culpable than someone who fails to check their own work.

What happens if a fake citation reaches the court?

The likely consequences include a wasted costs order, a referral to the SRA, disciplinary action and, in serious cases, contempt proceedings. The courts have already referred lawyers to their regulators for it. Reporting the error promptly, and reporting yourself to the SRA, can make a real difference to the outcome.

Do we have to tell clients we use AI?

There's no SRA rule requiring it. The Law Society suggests clear communication where appropriate. Decide your approach, put it in your terms of business or AI policy, and apply it consistently.

Do we need an AI policy?

The SRA doesn't require a document called an AI policy, but it does expect governance, systems and controls that cover AI. A short written policy is the simplest way to show you have them.

When does the competence statement change take effect?

The consultation closes on 3 December 2026. The SRA plans to publish the final statement in April 2027, with changes taking effect from September 2027.

Naz Keceli

CEO and Co-Founder

Published on the Khiliad Legal Insights series. For longer conversations on AI, technology, and legal practice, get in touch directly.

Not sure which way in?

A first conversation costs nothing. We'll tell you where we'd start, and if the answer is not yet, we'll say so.